fdbw commited on
Commit
c820e8b
·
verified ·
1 Parent(s): 2c38221

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +13 -9
app.py CHANGED
@@ -10,7 +10,7 @@ device = 0 if torch.cuda.is_available() else -1
10
  print(f"正在加载 Whisper 模型,使用设备: {'GPU' if device == 0 else 'CPU'}...")
11
  MODEL_NAME = "openai/whisper-small"
12
 
13
- # 加载 Pipeline 和 Processor(用于将文本 Prompt 转为 prompt_ids)
14
  pipe = pipeline(
15
  task="automatic-speech-recognition",
16
  model=MODEL_NAME,
@@ -18,13 +18,18 @@ pipe = pipeline(
18
  )
19
  processor = AutoProcessor.from_pretrained(MODEL_NAME)
20
 
21
- # 生成方言引导的 prompt_ids,解决 transformers 的 prompt 报错
22
  prompt_text = "以下是普通话及四川方言对话:"
23
- prompt_ids = processor.get_prompt_ids(prompt_text)
 
 
 
 
 
24
 
25
  def transcribe_stream(audio_chunk, state=""):
26
  """
27
- 处理音频流:专为中文及方言口音优化,解决 prompt 参数报错
28
  """
29
  if audio_chunk is None:
30
  return state, state
@@ -45,11 +50,10 @@ def transcribe_stream(audio_chunk, state=""):
45
  result = pipe(
46
  {"sampling_rate": sample_rate, "raw": audio_data},
47
  generate_kwargs={
48
- "language": "zh", # 强制锁死中文
49
- "task": "transcribe",
50
  "max_new_tokens": 128,
51
- "repetition_penalty": 1.2, # 防止乱码循环
52
- "prompt_ids": prompt_ids # 正确传入 Prompt Token IDs
53
  }
54
  )
55
 
@@ -61,7 +65,7 @@ def transcribe_stream(audio_chunk, state=""):
61
 
62
  return state, state
63
 
64
- # 构建 Gradio 6.x 流式 Web 界面
65
  with gr.Blocks(title="Whisper 实时字幕转写系统") as demo:
66
  gr.Markdown("# 🎙️ 实时语音字幕转写系统(中文/方言优化版)")
67
  gr.Markdown("点击 **`录制`** 后直接说话,文字将实时滚动出现。")
 
10
  print(f"正在加载 Whisper 模型,使用设备: {'GPU' if device == 0 else 'CPU'}...")
11
  MODEL_NAME = "openai/whisper-small"
12
 
13
+ # 加载 Pipeline 和 Processor
14
  pipe = pipeline(
15
  task="automatic-speech-recognition",
16
  model=MODEL_NAME,
 
18
  )
19
  processor = AutoProcessor.from_pretrained(MODEL_NAME)
20
 
21
+ # 1. 转换为 PyTorch Tensor 格式,彻底解决 TypeError
22
  prompt_text = "以下是普通话及四川方言对话:"
23
+ prompt_ids = processor.get_prompt_ids(prompt_text, return_tensors="pt")
24
+ if device == 0:
25
+ prompt_ids = prompt_ids.cuda()
26
+
27
+ # 2. 获取锁定中文的解码 ID
28
+ forced_decoder_ids = processor.get_decoder_prompt_ids(language="chinese", task="transcribe")
29
 
30
  def transcribe_stream(audio_chunk, state=""):
31
  """
32
+ 处理音频流:专为中文及方言口音优化,兼容 PyTorch 类型
33
  """
34
  if audio_chunk is None:
35
  return state, state
 
50
  result = pipe(
51
  {"sampling_rate": sample_rate, "raw": audio_data},
52
  generate_kwargs={
53
+ "forced_decoder_ids": forced_decoder_ids, # 强制锁定中文任务
54
+ "prompt_ids": prompt_ids, # 传入 Tensor 类型的 Prompt
55
  "max_new_tokens": 128,
56
+ "repetition_penalty": 1.2 # 防止乱码循环
 
57
  }
58
  )
59
 
 
65
 
66
  return state, state
67
 
68
+ # 构建 Gradio 界面
69
  with gr.Blocks(title="Whisper 实时字幕转写系统") as demo:
70
  gr.Markdown("# 🎙️ 实时语音字幕转写系统(中文/方言优化版)")
71
  gr.Markdown("点击 **`录制`** 后直接说话,文字将实时滚动出现。")